The present invention relates to computer-implemented methods and data structures for producing candidate parent entities that are ranked in accordance with ranking information associated with given child entities and, in particular, to such methods for use with software parsers and data dictionaries, for example, of the kind utilized in a system for automated reading, validation, and interpretation of hand print, machine print, and electronic data streams.
Optical character recognition (OCR) systems and digital image processing systems are known for use in automatic forms processing. These systems deal with three kinds of data: physical data, textual data, and logical data. Physical data may be pixels on a page or positional information related to those pixels. In general, physical data is not in a form to be effectively used by a computer for information processing. Physical data by itself has neither useable content nor meaning. Textual data is data in textual form. It may have a physical location associated with it. It occurs in, for example, ASCII strings. It has content but no meaning. We know what textual data says, but not what it means. Logical data has both content and meaning. It often has a name for what it is.
For example, there may be a region of black pixels in a certain location on an image. Both the value of the pixels and their location are physical data. It may be determined that those pixels, when properly passed through a recognizer, say: xe2x80x9c(425) 867-0700xe2x80x9d Content has been derived from the physical data to generate textual data. If we now know that text of this format (or possibly at this location on a preprinted form) is a telephone number, the textual data becomes logical data.
To facilitate reconciliation of imperfections in physical data and shortcomings of the recognition process, each recognized element of textual data, e.g., a character, may be represented by a ranked group of unique candidates called a xe2x80x9cpossibility set.xe2x80x9d A possibility set includes one or more candidate information pairs, each including a xe2x80x9cpossibilityxe2x80x9d and an associated confidence. In the context of an OCR system, the confidence is typically assigned as part of the recognition process. For computational efficiency, the confidences may be assigned within an appropriate base-2 range, e.g., 0 to 255, or a more compact range, such as 0 to 7. For example, FIG. 1 shows an enlarged view of an individual glyph 20 that may be physically embodied as a handwritten character or as a digital pixel image of the handwritten character. From glyph 20, an optical character recognition process may generate the possibility set shown in TABLE 1 by assigning possibilities and associated confidences:
FIG. 2 shows a series of sibling glyphs 22, which are known as xe2x80x9csiblingsxe2x80x9d because they share the same parent word 24. The sibling glyphs 22 can be represented by the four possibility sets as shown in the following TABLE 2:
The possibilities of these four possibility sets can be readily combined to form 36 unique strings: xe2x80x9cchorxe2x80x9d, xe2x80x9cohorxe2x80x9d, xe2x80x9cehorxe2x80x9d, xe2x80x9ccnorxe2x80x9d, xe2x80x9ccrorxe2x80x9d, etc. The number of unique strings is determined by the product of the number of character possibilities in each possibility set, i.e., 3xc3x973xc3x972xc3x972=36.
To gage or verify their accuracy, the unique xe2x80x9ccandidatexe2x80x9d strings may be processed by a xe2x80x9cdictionaryxe2x80x9d of valid outcomes. In the context of OCR, a dictionary is a filter. It has content and rules. Each candidate string processed by the dictionary is subject to one of three possible outcomes: it is passed, it is rejected, or it is modified into a similar string that passes. One example of a dictionary is based on the English language. For parent word 24 of FIG. 2, the candidate strings xe2x80x9cchorxe2x80x9d and xe2x80x9ceharxe2x80x9d would be rejected by such a dictionary, while xe2x80x9ccharxe2x80x9d would be passed.
Because dictionaries often have a very large amount of content against which a candidate string is compared, it may be unduly time-consuming to apply the dictionary to all possible strings. To improve efficiency, it is desirable, before applying a dictionary, to rank the candidate strings in order of some confidence based on the accuracy of recognition. In this way the candidate strings having the highest confidence of having been accurately recognized are processed by the dictionary first. Rules can then be used to determine when to stop dictionary processing, e.g., when enough candidate strings have been processed to have isolated the best candidate strings (with a certain probability). A convenient way to rank candidate strings is to calculate string confidences based on the confidences of the component character possibilities that make up each candidate string. A set of candidate strings and their associated string confidences is referred to as an xe2x80x9calt-set.xe2x80x9d
One way to rank parent candidates for creating an alt-set is to add the child confidences for each parent candidate. In the above example, xe2x80x9cchorxe2x80x9d would have a ranking of 615 (the sum of the confidences associated with the individual characters c-h-o-r), xe2x80x9cohorxe2x80x9d would have a ranking of 565, xe2x80x9cehorxe2x80x9d would have a ranking of 515, etc. Combining the possibility sets to form the 36 unique strings and to calculate their rankings is simple in this example. However, there is no obvious way to read the strings out in ranked order. The strings must first be assigned a ranking, then ordered or sorted based on their assigned rank. This ordering or sorting step becomes especially problematic for longer strings formed from sibling possibility sets having a greater number of possibilities. By way of illustration, a hypothetical 10-character parent word in which each child possibility set includes 10 possibilities would result in 10 billion unique strings. It would be a very time-consuming and computationally expensive task to rank and order 10 billion 10-character strings.
Another known way of improving the efficiency of dictionaries is to use specialized dictionaries that contain smaller amounts of content than a more generalized dictionary but that are limited in their application. One such specialized dictionary is an xe2x80x9cn-gramxe2x80x9d dictionary, which includes information about the frequency in which certain character sequences (e.g., two-letter, three-letter, etc.) occur in the English language. For example, the two-letter combination xe2x80x9cQuxe2x80x9d (a 2-gram) occurs in English words much more frequently than xe2x80x9cQo.xe2x80x9d To benefit from an n-gram dictionary, the confidence assigned to an n-gram is some combination of (1) the aggregate character confidences and (2) the n-gram frequency provided by the n-gram dictionary. Thus, recognition may have produced Oueen and Queen where the first character has the possibility set: poss=O, conf=200; poss=Q, conf=100, but in the English language xe2x80x9cQuxe2x80x9d happens much more often than xe2x80x9cOuxe2x80x9d, so the 2-gram dictionary would help determine that Queen is the more likely parent string.
A need exists for a method of generating candidate strings in ranked order on an as-needed basis and, more generally, for a method of generating ranked parent candidates on an on-demand basis from a series of sibling possibilities. A need also exists for such a method that can be used with data at different logical levels in a logical data hierarchy, such as n-grams, words, and phrases.
In accordance with the present invention, methods of organizing a series of sibling data entities are provided for preserving sibling ranking information associated with the sibling data entities and for attaching the sibling ranking information to a joint parent of the sibling data entities to facilitate on-demand generation of ranked parent candidates. A rollup function of the present invention builds a rollup matrix containing information about the sibling entities and the sibling ranking information and provides a method for reading out the ranked parent candidates from the rollup matrix in order of their parent confidences, which are based on the sibling ranking information. Parent confidences may also be based, in part, on n-gram ranking or other ranking information.
External to the rollup function of the present invention, sibling entities are generated and passed to the rollup function for processing. Generation of a series of sibling entities may, in the context of OCR, involve optical scanning, recognition processing, and parsing. Each sibling entity comprises one or more ranked child possibilities, each having an associated child confidence. The number of child possibilities in a sibling entity is referred to as the xe2x80x9cchild populationxe2x80x9d of the sibling entity. Each sibling entity may include a range of child confidences, one of which is the maximum child confidence.
In one aspect of the invention the rollup function is implemented in computer software operable on a digital computer. The rollup matrix is modeled as a three-dimensional data array called a rollup table. The rollup table serves as a convenient visual aid to understanding the nature of the rollup matrix and operation of the rollup function. What is the matrix? It should be understood that nothing in the foregoing description of the rollup table should be construed as limiting the scope of the invention to implementation of the rollup matrix in data arrays. Other data structures, such as linked lists, are also suitable for implementing the rollup function of the present invention. It should be understood, therefore, that the term xe2x80x9crollup matrixxe2x80x9d as used herein shall mean data tables, linked lists, and any other device for defining relationships between nodes in a data structure, where such nodes include one or more elements of data and one or more relationships to other nodes, procedures, or nested rollup functions. Furthermore, it will be apparent from the foregoing description of the invention that while the invention is suitable for use with OCR technology, it is also suitable for use with processing of other types of content-bearing data in which uncertainty in the data content is sought to be resolved. Non-OCR applications of the invention involving resolution of empirical uncertainty may include, for example, bioinformatics systems for analyzing gene sequencing information.
After receiving a series of sibling data entities, a matrix initialization routine of the rollup function establishes a rollup table and sizes it based on properties of the sibling entities. In particular, the rollup table is sized to include a series of xe2x80x9ccolumnsxe2x80x9d equal in number to the number of sibling entities received. The dimension of the rollup table spanned by the columns is referred to as the xe2x80x9cwidthxe2x80x9d of the table. The rollup table is sized in a xe2x80x9cheightxe2x80x9d dimension based on a number of xe2x80x9crows,xe2x80x9d with each having a row position indicating its position along the height dimension of the data table. The number of rows, and consequently the height of the table, is based on the sum of the maximum child confidences of the sibling entities. In practice, the number of rows may be established as equal to the sum of the maximum child confidences plus one. The rollup table is sized in a xe2x80x9cdepthxe2x80x9d dimension based on the largest of the child populations of the sibling entities. The rollup table is a collection of xe2x80x9cnodes,xe2x80x9d each located in the rollup table at a position defined by column, row position, and a depth position in the depth dimension.
Once the rollup function has established the rollup table, a loading routine of the rollup function then loads the sibling entities into the rollup table in a predetermined loading sequence beginning with loading a first sibling entity in a first column of the series of columns. Each sibling entity is loaded in sequence, from the first sibling entity to the last sibling entity in the series. If the sibling entities have no serial relationship, then an arbitrary, but ordered sequence of loading is chosen. Each child possibility of the first sibling entity is loaded into a node of the rollup table located at the first column and at the row having a row position corresponding to the child confidence of the child possibility being loaded. The rollup function then proceeds to load the second sibling entity in the series in a second column. For the second and each subsequent sibling entity and column, the rollup function loads each child possibility in one row of the current column for each row of the immediately preceding column having a filled node. The child possibilities of the second sibling entity are loaded in rows of the second column that have row positions offset from the row positions of filled nodes of the immediately preceding column (i.e., the first column) by an offset amount corresponding to the child confidence of the child possibility being loaded in the second column. The child possibilities of the third sibling entity are loaded in rows of the third column having row positions offset from the row positions of filled nodes of the second column by an offset amount corresponding to the child confidence of the child possibility being loaded in the third column, and so on, until the last sibling entity has been loaded in the last column of the rollup table. Each entry in the last column of the rollup table is a terminal element. Due to different confidence values that may be associated with multiple child possibilities of each of the sibling entities, the loading sequence may result in the loading of multiple elements in a particular column and row position. During loading, if a node has already been filled with a child possibility, the loading routine offsets in the depth of the rollup table until it reaches an unoccupied node, then fills that node.
Upon completion of the loading sequence, another aspect of the invention involves a roll-out routine of the rollup function, which may be used to read parent candidates from the rollup table according to their parent confidences. The reading of parent candidates, known as xe2x80x9croll-out,xe2x80x9d begins with a terminal element known as an entry point. Each parent candidate is assembled in a sequence opposite the sequence in which the rollup table was loaded, as follows: After reading a terminal element from the last column, the roll-out routine then reads a next-to-last element from the node located at a next-to-last column immediately preceding the last column and at a row position less than the row position of the entry point by an amount equal to the child confidence associated with the terminal element. The next-to-last element is then prepended to the terminal element to form a string tail. A prefix element is read from a node located in the column immediately preceding the next-to-last column and at a row position less than the node of the next-to-last element by an amount equal to the confidence of the next-to-last element. The prefix element is then prepended to the string tail. If the sibling entities forming the rollup table have no serial relationship, then prepending involves combining the elements in reverse order of their loading in the rollup table. This reading process is repeated until the roll-out routine reaches the first column, completing roll-out of the parent candidate. If more than one element is located at a particular column and row location (i.e., elements are stored at more than one depth position), then the roll-out routine will continue reading parent candidates beginning from the same entry point until elements at all occupied nodes at all depths in the appropriate columns and rows have been read and all parent candidates having the same parent confidence have been rolled out, or until the desired number of parent candidates have been rolled out. The roll-out process is merely repeated for further parent candidates.
The method of loading the data table dictates that each row position corresponds to the parent rank of each parent candidate assembled from a terminal element located at that row position. The parent candidate (or candidates) with the greatest parent confidence may be read from the rollup matrix by beginning at a maximal node located at the last column and at the row of greatest row position. Consequently, parent candidates may be read in decreasing order of parent rank by merely assembling parent candidates in sequence, beginning with terminal element(s) located at the maximal node and continuing to read from the rollup table at entry points of decreasing row position until all parent candidates have been assembled. The process of building a rollup matrix and rolling-out parent candidates to form alt-sets can be repeated at each level in the data hierarchy. If desired, rollup functions can be nested by storing a nested xe2x80x9cchildxe2x80x9d rollup function pointer at a node of a parent roll-up table.
Given the foregoing description of the invention, the use of software counters to facilitate the loading of the rollup matrix and the roll-out of parent candidates will be understood by those skilled in the art.
In another aspect of the invention, the rollup matrix is established in a computer memory using a plurality of memory pointers in place of the 3-dimensional data array of the rollup table. In this aspect of the invention, the terms xe2x80x9crowsxe2x80x9d and xe2x80x9ccolumnsxe2x80x9d are arbitrary but are used herein to denote memory locations within the rollup matrix. In reality, each node of the rollup matrix includes a pointer to other nodes which contain a child possibility of an adjacent sibling entity. If a node must point to more than one child possibility, as in the case of multiple child possibilities at a particular column and row position, the node will include multiple pointers. When these multi-pointer nodes are encountered by the roll-out routine, a branch is indicated so that all pointers of each node are followed before moving to the next entry point.
Nodes occupying entry points shall be referred to as xe2x80x9centry nodes.xe2x80x9d Entry nodes further include a parent confidence which the roll-out routine recognizes as assigned to the parent candidate assembled beginning with the entry node. Entry nodes may also include a pointer to the next entry node in the matrix, which may have the same parent confidence or a lesser parent confidence. Nodes in the xe2x80x9cfirst column,xe2x80x9d loaded with a child possibility of the first possibility set, may include a return pointer that may direct the roll-out routine to output the completed parent candidate for verification (e.g., using a dictionary) or to proceed to the next entry node for generation of the next parent candidate. Nodes at any location in the rollup matrix may also include a pointer to an entry node of a nested rollup matrix.
In yet another aspect of the invention, n-gram possibility sets are generated using an n-gram rollup function in accordance with the present invention. Comparison of parent candidate n-grams against an n-gram dictionary allows n-gram candidates to be weighted in accordance with their relative frequencies of occurrence in the context of, for example, the English language. Possibility sets including n-grams are readily accommodated in establishing the rollup matrix. For 3-grams, the nodes are loaded with the 3-grams at a row position which is the aggregate of the confidence of the central character (of the 3-gram) and the dictionary-provided frequency of the 3-gram. In this aspect of the invention, child possibilities in the first and last columns of the rollup matrix must be prepended and appended, respectively, with nulls (or spaces) so that all child possibilities are 3-grams. Further, the 3-gram child possibilities must be loaded in the rollup matrix so that when the parent candidates are rolled-out, all adjacent 3-grams assembled in a parent candidate share two characters. For example, xe2x80x9coutxe2x80x9d in the first column will fit with xe2x80x9cutsxe2x80x9d in the second column, but not with xe2x80x9cnts.xe2x80x9d
In the context of OCR, the rollup function of the present invention is useful at every level of textual hierarchy. Rollup functions also avoid fatal problems often encountered by prior art string generators, which create strings from a series of possibility sets. Existing string generators suffer from three major problems. First, they are combinatorically expensive in memory usexe2x80x94needing a place in memory for each possible string. Second, string generators must trim strings before generating all possible strings because of limited space to store the combinatorically-many strings. Therefore, it is possible for string generators to result in higher-confidence strings being abandoned while lower-confidence strings are preserved. Third, string generators do not guarantee that strings of the same confidence, once ordered, retain that order.
The present invention gets around all these-problems in a natural way. First, the rollup function is only geometrically expensive of memory, not combinatorically. Tables generated by prior art systems grow as Lxc3x97nL, where n is the number of possibilities per possibility set and L is the number of possibility sets (i.e., the string length). There are nL strings of length L that can be generated. By comparison, the rollup matrix of the present invention grows as 2xc3x97CFmaxxc3x97L2, where CFmax is the highest confidence value in any possibility set. A significant savings over prior art systems. For L=10, n=3, and CFmax=20, and allowing 1 byte per ASCII character, approximately 590,490 bytes would be required for ranking tables of prior art systems; while only 12,000 bytes are required for the rollup matrixxe2x80x94a savings of 98%. Second, candidate strings can be read out of a rollup table in their decreasing order of confidence without having to store unneeded strings in memory, while never skipping a higher-confidence parent candidate for a lower confidence one. The rollup matrix does not change size with the number of generated strings. Therefore, all strings are preserved and there is no trimming of strings ever required. Third, no reordering of parent strings ever takes place because the rollup matrix is unchanging. Consequently, strings of the same confidence remain in their original order.
Parent candidates can be read from the rollup matrix in decreasing or increasing order of parent confidence. First, a parent candidate having a desired confidence value can easily be selected from the matrix by a confidence stored in association with an entry node of the parent candidate. Parent candidates having lesser (or greater) confidences can then be read until a desired lesser (or greater) confidence level is reached. This process can be repeated until a predetermined number of parent candidates have been obtained or until all possible parent candidates have been rolled-out. The rollup function can be interrupted while reading out a parent candidate to handle some other process, such as verifying the most recently rolled-out parent candidate using a dictionary. The rollup function easily returns to where it left off in the rollup matrix to read out the next-ranked parent candidate by returning to the location in the rollup matrix that was being accessed when the interruption occurred. The rollup function of the present invention provides the above-described benefits without requiring the production of all of the parent candidates before subsequent ranking. If a particular child possibility occurs with at most one confidence value in a possibility set, then the last rolled-out string is the pointer structure. Even in the case of allowed duplication, returning to the rollup function is as simple as storing a pointer to the next entry point in the rollup matrix and storing a pointer to each position of the table, which may be accomplished by freezing the internal pointer structure.
The rollup function of the present invention is, of course, not limited to strings. Any parent entity can receive rollup-produced alt-sets from its child entities. For example, gene sequence information prepared from a human, an animal, a plant, or any other living organism may be parsed into its nucleotides, each of which may be represented by an alt-set. Sibling nucleotide alt-sets can then be loaded into a rollup matrix for the parent gene. In this way, the frequency of naturally-occurring nucleotide and coding sequence variations can easily be represented by the child confidences associated with child possibilities of each alt-set. Inaccuracies inherent in the gene sequencing process can be similarly represented by the child confidences.